Enrichment Factor

EF measures the concentration of true active compounds in a selected top fraction (χ%) of a ranked list relative to a random distribution. A value greater than 1 indicates better-than-random performance.

• Formula: EF = (Hits_selected / N_selected) / (Hits_total / N_total)
• Common Thresholds: EF at 1% (EF1%) and 5% (EF5%) are standard benchmarks
• Interpretation: An EF1% of 20 means actives are 20x more concentrated in the top 1% than by chance

The maximum possible EF value is fundamentally constrained by the overall hit rate of the screened library. A library with very few actives can yield a high EF, while a dense library cannot.

• Max EF: EF_max = 1 / χ (when χ ≤ hit rate) or N_total / (χ * N_total) for sparse libraries
• Practical Limit: For a 1% selection, the theoretical maximum EF is 100
• Comparison Caveat: EF values from libraries with different hit rates cannot be directly compared without normalization

EF is intentionally biased toward early recognition, making it insensitive to the ranking quality of actives appearing beyond the chosen cutoff threshold. This is both a strength and a limitation.

• Strength: Aligns with practical drug discovery, where only the top-ranked compounds are physically tested
• Limitation: Two models with identical EF1% may have vastly different overall ranking quality
• Complementary Metric: Pair EF with AUC-ROC or Boltzmann-Enhanced Discrimination of ROC (BEDROC) for a holistic view

Determining whether an observed EF value is statistically meaningful requires comparing it against a null distribution generated by random ranking. This prevents over-interpretation of chance fluctuations.

• Null Hypothesis: The ranking is random; active compounds are uniformly distributed
• Method: Generate 10,000+ random rankings to build an empirical null distribution of EF values
• p-value: The fraction of random trials achieving an EF ≥ the observed EF
• Confidence Intervals: Report 95% confidence intervals alongside point estimates of EF

EF is the primary metric for validating a virtual screening protocol using known actives and decoys before prospective application. It answers: 'If we had used this model, would we have found the known drugs?'

• Benchmark Datasets: DUD-E (Directory of Useful Decoys, Enhanced) and DEKOIS provide standardized ligand sets
• Decoy Selection: Physico-chemical property-matched decoys are critical; biased decoys inflate EF artificially
• Leave-One-Out Cross-Validation: Ensures the model generalizes beyond its training set when calculating EF

While related to the Receiver Operating Characteristic curve, EF provides a discrete, threshold-specific snapshot rather than a global measure. Understanding the distinction is crucial for correct reporting.

• ROC AUC: Integrates performance across all possible thresholds; a global metric
• EF: Measures performance at a single, pre-defined early threshold; a local metric
• ROC Enrichment: The true positive rate at a specific false positive rate, directly proportional to EF (ROC Enrichment = EF * χ)
• Best Practice: Report both EF (at 1%, 5%) and AUC-ROC for comprehensive model evaluation

The Area Under the Receiver Operating Characteristic Curve is a threshold-independent metric evaluating a binary classifier's ability to distinguish active from inactive compounds. Unlike the Enrichment Factor, which focuses on early recognition, AUC-ROC assesses global ranking quality across all thresholds.

• Plots True Positive Rate vs. False Positive Rate
• A score of 1.0 indicates perfect separation
• A score of 0.5 is equivalent to random selection
• Complements EF by measuring overall model calibration

A mathematical function used in molecular docking to approximate the binding free energy of a protein-ligand complex. It enables the rapid ranking of different binding poses and compounds, directly generating the ranked list from which the Enrichment Factor is calculated.

• Force-field-based: Uses physics equations (van der Waals, electrostatics)
• Empirical: Weights terms like hydrogen bonds and hydrophobic contacts
• Knowledge-based: Derives potentials from statistical analysis of known structures
• The accuracy of the scoring function fundamentally limits the maximum achievable EF

A metric for comparing the similarity of two sets, most commonly applied to binary molecular fingerprints. It quantifies the structural overlap between two chemical compounds and is often used as a baseline method against which more sophisticated machine learning models are benchmarked using the Enrichment Factor.

• Defined as: c / (a + b - c) where c is the intersection
• Ranges from 0 (no shared features) to 1 (identical fingerprints)
• A simple Tanimoto-based virtual screen provides a baseline EF
• GNNs and deep learning aim to surpass this structural similarity ceiling

A standard quantitative measure of the average distance between atoms of superimposed protein structures or docked ligand poses. It is the primary metric for assessing pose prediction accuracy, which is a prerequisite for a meaningful Enrichment Factor calculation.

• Typically reported in Ångströms (Å)
• A threshold of ≤ 2.0 Å is commonly considered a successful pose
• If the top-ranked pose has high RMSD, the scoring function failed
• Validating docking protocol via RMSD is essential before calculating EF

The computational process of generating a diverse ensemble of low-energy three-dimensional shapes that a flexible ligand molecule can adopt. Inadequate sampling leads to missed binding modes, artificially depressing the Enrichment Factor by failing to identify true actives.

• Systematic search: Rotates bonds in discrete increments
• Stochastic search: Randomly changes torsion angles (e.g., Monte Carlo)
• Molecular Dynamics: Simulates physical motion over time
• The EF is sensitive to both the thoroughness and speed of this search

A computational technique used to rapidly evaluate large chemical libraries to identify molecules most likely to bind to a drug target. The Enrichment Factor is the primary metric for quantifying the success of a virtual screening campaign.

• Structure-Based (SBVS): Uses the 3D structure of the target protein
• Ligand-Based (LBVS): Uses known active compounds to find similar ones
• A typical screen ranks millions of compounds to select a few hundred for testing
• EF directly answers: 'How much better than random is this screen?'